Experiments on Kaldi-Based Forced Phonetic Alignment for Brazilian Portuguese

Abstract

Forced phonetic alignment (FPA) is the task of associating a given phonetic unit to a timestamp interval in the speech waveform. Phoneticians are able mark the boundaries with precision, but as the corpus grows it becomes infeasible to do it by hand. For Brazilian Portuguese (BP) in particular, only three tools appear to perform FPA: EasyAlign, Montreal Forced Aligner (MFA), and UFPAlign. Therefore, this work aims to develop resources based on Kaldi toolkit for UFPAlign, including their release alongside all scripts under open licenses; and to bring forth a comparison to the other two aforementioned aligners. Evaluation took place in terms of the phone boundary metric over a dataset of 385 hand-aligned utterances, and results show that Kaldi-based aligners perform better overall, and that UFPAlign models are more accurate than MFA’s. Furthermore, complex deep-learning-based approaches did not seem to improve performance compared to simpler models.